Elucidating the performance and mechanisms of membrane separation: the use of artificial intelligence and a case study of produced water treatment
Date
2023
Authors
Jeong, Nohyeong, author
Tong, Tiezheng, advisor
Carlson, Kenneth, committee member
Sharvelle, Sybil, committee member
Bandhauer, Todd, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Pressure-driven membrane technologies such as nanoļ¬ltration (NF) and reverse osmosis (RO) have been widely used in water and wastewater treatment because of their effective removal of contaminants and exceptional energy efficiencies. The performance of NF and RO membranes is regulated by the well-documented permeability-selectivity tradeoff, in which an increase of membrane permeability typically occurs at the expense of membrane selectivity and vice versa. To break the upper bound of this tradeoff and further enhance the efficiency of NF and RO treatment, a mechanistic understanding of the solute transport across membranes with pore sizes at the nanometer- or angstrom-scale is required. Current theoretical models relating to solute transport across membranes are limited as the models require precise acquisition of multiple parameters. Machine learning (ML) models, a data-driven approach, have been applied to predict membrane performance and elucidate the membrane separation mechanisms. However, whether the ML models possess appropriate knowledge on membrane separation mechanisms has not yet been studied. Probing knowledge of ML models on membrane separation mechanisms can enhance the reliability of the ML model, which is of great importance to the implementation of ML models for decision-making processes, such as membrane design and selection. Moreover, contrary to the well-controlled experiments for studying the mechanisms or models associated with solute transport, where a limited number of defined solutes are present, membrane treatment has been used to treat wastewater containing diverse organic and inorganic compounds. Thus, along with fundamental research on predictive ML models for membrane performance, investigating the performance of membranes for treating wastewater with complex compositions is also valuable to provide knowledge of solute transport across membranes in practical applications. In this thesis, I present both a fundamental study of probing solute transport across NF and RO membranes using ML models and an applied study that explores membrane treatment of unconventional oil and gas (UOG) produced water. First, the reliability of the ML model as a tool to predict membrane performance was investigated. Specifically, the influence of data leakage on the ML model performance, as well as the solution to prevent this issue, was explored to evaluate the prediction capability of the ML model objectively. I discovered that data leakage can lead to falsely high prediction accuracy of the ML model, and appropriate data splitting for the training, validation, and testing dataset is necessary to avoid data leakage. Second, the underlying knowledge of ML models for organic and inorganic solute transport across polyamide membranes was investigated by using a model interpretation method (i.e., Shapley additive explanation, SHAP). I not only tested whether ML models are able to possess adequate knowledge on solute transport, but also utilized the SHAP method to reveal solute transport mechanisms that are typically obtained using tedious, well-controlled experiments. For the ML model applied to predicting the rejection of organic constituents by NF and RO membranes, I found that the ML model had proper knowledge of size exclusion, but its understanding of electrostatic interaction and adsorption remains rudimentary. By using ML to predict the rejection of inorganic constituents, I elucidated that explainable artificial intelligence (XAI) can capture the major governing mechanisms of ion/salt transport across polyamide membranes (i.e., size exclusion and electrostatic interaction), which have different importance for the transport of single salt, cation, anion transport in mixture salt solution. Lastly, the performance of RO/NF membranes for the treatment of UOG produced water was explored as a case study, which comprehensively investigated the chemical composition and toxicity level of the treated water. NF permeates, which still had high salinities and high boron concentrations, were found to be inappropriate for irrigation and livestock drinking water, while RO membranes effectively removed most pollutants and met most water quality standards for beneficial reuse (i.e., irrigation and livestock drinking water). However, the chloride concentrations and sodium adsorption ratio (SAR) values of RO permeates were still higher than the recommended thresholds for irrigation. Also, surfactants with molecular weights higher than the molecular weight cut-off of RO/NF membranes were able to traverse through the membrane, indicating that NF and RO are not complete barriers against organic contaminants. The toxicity test results of NF and RO permeates demonstrated that NF permeates were still toxic to Daphnia, while RO permeates showed less toxicity than NF permeates or no toxicity. The toxicity level of NF and RO permeates showed a correlation with salinity in the permeates, which might be the main driver of the toxicity. I envision that my thesis provides a framework to evaluate the knowledge and reliability of ML model predictions, while presenting a comprehensive investigation on membrane performance and the potential risks associated with membrane treatment of UOG produced water for beneficial reuse. The knowledge gained in this thesis improves our capability for rational membrane material design and selection, which has the potential to lead to more efficient NF and RO technologies for sustainable water and wastewater treatment.